基于低剂量CT图像序列的三维肺实质提取

发布时间：2018-01-09 00:29

本文关键词：基于低剂量CT图像序列的三维肺实质提取　出处：《郑州大学》2017年硕士论文　论文类型：学位论文

【摘要】：肺癌是患病率和病死率最高的恶性肿瘤。随着近年空气污染和雾霾越来越严重,开展肺癌的早期筛查愈来愈紧迫。与X线胸片相比,采用低剂量CT(low-dose computed tomography,LDCT)对肺癌高危人群进行筛查可使肺癌病死率下降20%。由此可以预见,基于低剂量螺旋CT取代X线胸片进行肺部疾病筛查是未来发展的趋势。肺癌早期一般以肺结节的形式出现,而肺实质分割是肺结节检测的重要前提。鉴于此,本文基于低剂量CT图像序列,研究精准的三维肺实质提取算法。本文的主要工作如下:(1)精准的二维肺实质分割算法针对噪声、肺部区域的不均匀性以及胸膜与肺结节的粘连会影响二维肺实质分割精度的问题,提出精准的二维肺实质分割算法。首先,采用传统的肺实质分割方法,经过预处理、二值化、去除气管和支气管以及左右肺分离步骤,初步分割出肺实质区域;然后针对肺实质初步分割中近胸膜结节与血管易被错误排除在肺实质区域之外的问题,提出了一种基于陡变度的肺实质边缘修补算法,通过检测肺实质边缘的陡变点、提取肺实质边缘的拐角点,以及选取并连接重要拐角点对,即可准确修补肺实质边缘凹陷,进而得到完整的肺实质图像。该算法能精确检测肺实质边缘的拐角点,进而能高效修补肺实质边缘凹陷。与文献法相比,提出算法的二维肺实质分割精度提升了0.76%,但分割速度较低,需进一步改进。(2)快速的低剂量CT图像序列自动化分割针对循环处理低剂量CT图像序列中的二维切片需消耗大量时间与人力的问题,提出一种改进的基于两级队列的3D并行区域生长方法,以快速实现图像序列分割的自动化。该算法在传统区域生长法的基础上,采用两级队列进行生长,并优先搜索边缘邻域点,加快了生长速度;同时详细考虑了种子点在相邻切片之间的变化情况,进而可实现对图像序列的自动化分割,并避免了分割错误。与传统区域生长法相比,提出算法的单幅图像平均处理时间减少了0.35s,体积重叠率提升了0.15%,过分割率降低了0.07%,欠分割率降低了0.04%。与文献法相比,体积重叠率提升了1.03%,过分割率降低了0.08%,欠分割率降低了1.2%,近胸膜结节包含率高达100%,单幅图像平均处理时间下降了0.09s。由此证明,提出算法能高效实现低剂量CT图像序列分割的自动化,为三维肺实质分割提供了便利条件。(3)三维肺实质提取最后,采用本文方法对多组低剂量CT图像序列进行处理,并对分割出的肺实质图像序列进行三维重建,以提取三维肺实质图像。实验结果表明,本文方法不仅能快速分割出高精度的肺实质图像序列,并且具有很好的三维可视化效果。
[Abstract]:Lung cancer is the malignant tumor with the highest morbidity and mortality. With the air pollution and haze becoming more and more serious in recent years, it is more and more urgent to carry out early screening of lung cancer. Low dose CT(low-dose computed tomography was used. LDCT screening of high risk groups of lung cancer can reduce the mortality of lung cancer by 20%, which can be predicted. Lung disease screening based on low dose spiral CT instead of chest radiography is a trend in the future. Lung cancer usually occurs in the form of pulmonary nodules in the early stage, and pulmonary parenchyma segmentation is an important prerequisite for the detection of pulmonary nodules. Based on the low dose CT image sequence, this paper studies a precise three-dimensional lung parenchyma extraction algorithm. The main work of this paper is as follows: 1) the precise two-dimensional lung parenchyma segmentation algorithm is aimed at noise. The inhomogeneity of lung region and the adhesion between pleura and pulmonary nodules will affect the segmentation accuracy of two-dimensional lung parenchyma. A precise two-dimensional segmentation algorithm of lung parenchyma is proposed. Firstly, the traditional method of lung parenchyma segmentation is adopted. After pretreatment, binarization, trachea and bronchus removal, and left and right lung separation steps, the lung parenchyma area was preliminarily separated. Then, aiming at the problem that the near pleural nodules and blood vessels are easily misruled out of the lung parenchyma region in the primary segmentation of pulmonary parenchyma, an algorithm of pulmonary parenchyma edge repair based on the degree of steepness is proposed. By detecting the sharp change point of the pulmonary parenchyma edge, extracting the corner point of the pulmonary parenchyma edge, and selecting and connecting the important corner pair, the indentation of the pulmonary parenchyma edge can be repaired accurately. The algorithm can accurately detect the corner point of the pulmonary parenchyma edge, and then it can effectively repair the indentation of the pulmonary parenchyma edge, compared with the literature method. The segmentation accuracy of two-dimensional lung parenchyma is improved by 0.76, but the segmentation speed is low. The fast automatic segmentation of low dose CT image sequence is needed to solve the problem that it takes a lot of time and manpower to process 2D slice in low dose CT image sequence. An improved 3D parallel region growth method based on two-level queue is proposed to automate image sequence segmentation. The algorithm is based on the traditional region growth method and uses two-level queue to grow. And priority search edge neighborhood points, accelerate the growth rate; At the same time, the variation of seed points between adjacent slices is considered in detail, which can realize the automatic segmentation of image sequences and avoid segmentation errors, compared with the traditional region growth method. The proposed algorithm reduces the average processing time of a single image by 0.35s, increases the volume overlap rate by 0.15s, and reduces the over-segmentation rate by 0.07%. Compared with the literature method, the volume overlap rate increased by 1.033%, the over-segmentation rate decreased by 0.08%, and the under-segmentation rate decreased by 1.2%. The inclusion rate of near-pleural nodules is as high as 100 and the average processing time of a single image is decreased by 0.09s. It is proved that the proposed algorithm can efficiently automate the segmentation of low-dose CT images. Three dimensional lung parenchyma segmentation provides a convenient condition for three-dimensional lung parenchyma extraction. Finally, we use this method to process multi-group low-dose CT image sequence. The segmented lung parenchyma image sequence is reconstructed to extract the three-dimensional lung parenchyma image. The experimental results show that the proposed method can not only segment the high-precision lung parenchyma image sequence quickly. And has the very good three-dimensional visualization effect.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R734.2;TP391.41

【相似文献】